MST

TensorFlow code for our ICCV 2019 paper "Multimodal Style Transfer via Graph Cuts"

Multimodal Style Transfer via Graph Cuts

This repository is for MST introduced in the following paper

Yulun Zhang, Chen Fang, Yilin Wang, Zhaowen Wang, Zhe Lin, Yun Fu, and Jimei Yang, "Multimodal Style Transfer via Graph Cuts", ICCV 2019, [arXiv]

The code is tested on Ubuntu 16.04 environment (Python3.6, TensorFlow1.2, scikit-learn 0.19.1, PyMaxflow 1.2.11, CUDA9.0, cuDNN5.1) with Titan X/1080Ti/Xp GPUs.

For K-Means clustering, we use sklearn package (scikit-learn 0.19.1). For Graph-cuts, we use PyMaxflow package (PyMaxflow 1.2.12). More enviroment information are available at src/environments.md.

PyTorch code for MST is on the way.

Test

# single content and style image pair
python test_MST.py --content='data/content/brad_pitt.jpg' --style='data/style/sketch.jpg'
# content and style image dir
python test_MST.py --content_dir='data/content' --style_dir='data/style'

Results

The test data (29 content images, 61 style images) are available at GoogleDrive and BaiduYun. The visual results (29x61 stylized images) are available at GoogleDrive and BaiduYun. The results are produced by using the model and parameters (e.g., alpha=1.0, K=3) of the paper.

Citation

If you find the code/results helpful in your resarch or work, please cite the following papers.

@InProceedings{zhang2019multimodal,
  author = {Zhang, Yulun and Fang, Chen and Wang, Yilin and Wang, Zhaowen and Lin, Zhe and Fu, Yun and Yang, Jimei},
  title = {Multimodal Style Transfer via Graph Cuts},
  booktitle = {ICCV},
  year = {2019}
}